git.sesse.net Git - ffmpeg/blob - libavcodec/x86/h264_i386.h

   1 /*
   2  * H.26L/H.264/AVC/JVT/14496-10/... encoder/decoder
   3  * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
   4  *
   5  * This file is part of FFmpeg.
   6  *
   7  * FFmpeg is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; either
  10  * version 2.1 of the License, or (at your option) any later version.
  11  *
  12  * FFmpeg is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with FFmpeg; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 /**
  23  * @file
  24  * H.264 / AVC / MPEG4 part10 codec.
  25  * non-MMX i386-specific optimizations for H.264
  26  * @author Michael Niedermayer <michaelni@gmx.at>
  27  */
  28
  29 #ifndef AVCODEC_X86_H264_I386_H
  30 #define AVCODEC_X86_H264_I386_H
  31
  32 #include <stddef.h>
  33
  34 #include "libavcodec/cabac.h"
  35 #include "cabac.h"
  36
  37 //FIXME use some macros to avoid duplicating get_cabac (cannot be done yet
  38 //as that would make optimization work hard)
  39 #if HAVE_7REGS
  40 static int decode_significance_x86(CABACContext *c, int max_coeff,
  41                                    uint8_t *significant_coeff_ctx_base,
  42                                    int *index, x86_reg last_off){
  43     void *end= significant_coeff_ctx_base + max_coeff - 1;
  44     int minusstart= -(intptr_t)significant_coeff_ctx_base;
  45     int minusindex= 4-(intptr_t)index;
  46     int bit;
  47     x86_reg coeff_count;
  48
  49 #ifdef BROKEN_RELOCATIONS
  50     void *tables;
  51
  52     __asm__ volatile(
  53         "lea   "MANGLE(ff_h264_cabac_tables)", %0      \n\t"
  54         : "=&r"(tables)
  55     );
  56 #endif
  57
  58     __asm__ volatile(
  59         "3:                                     \n\t"
  60
  61         BRANCHLESS_GET_CABAC("%4", "%q4", "(%1)", "%3", "%w3",
  62                              "%5", "%q5", "%k0", "%b0",
  63                              "%a11(%6)", "%a12(%6)", "%a13", "%a14", "%a15", "%16")
  64
  65         "test $1, %4                            \n\t"
  66         " jz 4f                                 \n\t"
  67         "add  %10, %1                           \n\t"
  68
  69         BRANCHLESS_GET_CABAC("%4", "%q4", "(%1)", "%3", "%w3",
  70                              "%5", "%q5", "%k0", "%b0",
  71                              "%a11(%6)", "%a12(%6)", "%a13", "%a14", "%a15", "%16")
  72
  73         "sub  %10, %1                           \n\t"
  74         "mov  %2, %0                            \n\t"
  75         "movl %7, %%ecx                         \n\t"
  76         "add  %1, %%"REG_c"                     \n\t"
  77         "movl %%ecx, (%0)                       \n\t"
  78
  79         "test $1, %4                            \n\t"
  80         " jnz 5f                                \n\t"
  81
  82         "add"OPSIZE"  $4, %2                    \n\t"
  83
  84         "4:                                     \n\t"
  85         "add  $1, %1                            \n\t"
  86         "cmp  %8, %1                            \n\t"
  87         " jb 3b                                 \n\t"
  88         "mov  %2, %0                            \n\t"
  89         "movl %7, %%ecx                         \n\t"
  90         "add  %1, %%"REG_c"                     \n\t"
  91         "movl %%ecx, (%0)                       \n\t"
  92         "5:                                     \n\t"
  93         "add  %9, %k0                           \n\t"
  94         "shr $2, %k0                            \n\t"
  95         : "=&q"(coeff_count), "+r"(significant_coeff_ctx_base), "+m"(index),
  96           "+&r"(c->low), "=&r"(bit), "+&r"(c->range)
  97         : "r"(c), "m"(minusstart), "m"(end), "m"(minusindex), "m"(last_off),
  98           "i"(offsetof(CABACContext, bytestream)),
  99           "i"(offsetof(CABACContext, bytestream_end)),
 100           "i"(H264_NORM_SHIFT_OFFSET),
 101           "i"(H264_LPS_RANGE_OFFSET),
 102           "i"(H264_MLPS_STATE_OFFSET) TABLES_ARG
 103         : "%"REG_c, "memory"
 104     );
 105     return coeff_count;
 106 }
 107
 108 static int decode_significance_8x8_x86(CABACContext *c,
 109                                        uint8_t *significant_coeff_ctx_base,
 110                                        int *index, uint8_t *last_coeff_ctx_base, const uint8_t *sig_off){
 111     int minusindex= 4-(intptr_t)index;
 112     int bit;
 113     x86_reg coeff_count;
 114     x86_reg last=0;
 115     x86_reg state;
 116
 117 #ifdef BROKEN_RELOCATIONS
 118     void *tables;
 119
 120     __asm__ volatile(
 121         "lea    "MANGLE(ff_h264_cabac_tables)", %0      \n\t"
 122         : "=&r"(tables)
 123     );
 124 #endif
 125
 126     __asm__ volatile(
 127         "mov %1, %6                             \n\t"
 128         "3:                                     \n\t"
 129
 130         "mov %10, %0                            \n\t"
 131         "movzbl (%0, %6), %k6                   \n\t"
 132         "add %9, %6                             \n\t"
 133
 134         BRANCHLESS_GET_CABAC("%4", "%q4", "(%6)", "%3", "%w3",
 135                              "%5", "%q5", "%k0", "%b0",
 136                              "%a12(%7)", "%a13(%7)", "%a14", "%a15", "%a16", "%18")
 137
 138         "mov %1, %k6                            \n\t"
 139         "test $1, %4                            \n\t"
 140         " jz 4f                                 \n\t"
 141
 142 #ifdef BROKEN_RELOCATIONS
 143         "movzbl %a17(%18, %q6), %k6\n\t"
 144 #else
 145         "movzbl "MANGLE(ff_h264_cabac_tables)"+%a17(%k6), %k6\n\t"
 146 #endif
 147         "add %11, %6                            \n\t"
 148
 149         BRANCHLESS_GET_CABAC("%4", "%q4", "(%6)", "%3", "%w3",
 150                              "%5", "%q5", "%k0", "%b0",
 151                              "%a12(%7)", "%a13(%7)", "%a14", "%a15", "%a16", "%18")
 152
 153         "mov %2, %0                             \n\t"
 154         "mov %1, %k6                            \n\t"
 155         "movl %k6, (%0)                         \n\t"
 156
 157         "test $1, %4                            \n\t"
 158         " jnz 5f                                \n\t"
 159
 160         "add"OPSIZE"  $4, %2                    \n\t"
 161
 162         "4:                                     \n\t"
 163         "addl $1, %k6                           \n\t"
 164         "mov %k6, %1                            \n\t"
 165         "cmpl $63, %k6                          \n\t"
 166         " jb 3b                                 \n\t"
 167         "mov %2, %0                             \n\t"
 168         "movl %k6, (%0)                         \n\t"
 169         "5:                                     \n\t"
 170         "addl %8, %k0                           \n\t"
 171         "shr $2, %k0                            \n\t"
 172         : "=&q"(coeff_count), "+m"(last), "+m"(index), "+&r"(c->low),
 173           "=&r"(bit), "+&r"(c->range), "=&r"(state)
 174         : "r"(c), "m"(minusindex), "m"(significant_coeff_ctx_base),
 175           "m"(sig_off), "m"(last_coeff_ctx_base),
 176           "i"(offsetof(CABACContext, bytestream)),
 177           "i"(offsetof(CABACContext, bytestream_end)),
 178           "i"(H264_NORM_SHIFT_OFFSET),
 179           "i"(H264_LPS_RANGE_OFFSET),
 180           "i"(H264_MLPS_STATE_OFFSET),
 181           "i"(H264_LAST_COEFF_FLAG_OFFSET_8x8_OFFSET) TABLES_ARG
 182         : "%"REG_c, "memory"
 183     );
 184     return coeff_count;
 185 }
 186 #endif /* HAVE_7REGS && !defined(BROKEN_RELOCATIONS) */
 187
 188 #endif /* AVCODEC_X86_H264_I386_H */